convez.ai
SkillAutomaçãoAvançado
0xadvait

AI Video Skill — Gerador com Aprendizado Contínuo

0xadvait/ai-video-skill

Geração de vídeo IA que aprende e se auto-melhora a cada clip

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill Python para Claude Code que gera vídeos end-to-end através de 6 modelos (Seedance 2.0, Kling, Wan, Veo, OmniHuman) com loop automático de controle de qualidade. A ferramenta não apenas cria clips, mas avalia cada geração, aprende com os erros e melhora seus prompts iterativamente.

Instalação

$npx convez-ai 0xadvait/ai-video-skill

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
10
Forks
1
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
v0.1.0
Issues abertas
0
Contribuidores
2

Casos de uso

  • Produção de vídeos comerciais e marketing com qualidade controlada automaticamente
  • Criação iterativa de conteúdo visual que se refina a cada geração
  • Automação de pipeline completo: frame inicial → prompt → síntese de voz → montagem → upscale

Veredito

Ferramenta robusta que vai além da geração simples: implementa QC automático e aprendizado iterativo. Exige familiaridade com APIs de IA de vídeo e Python avançado, mas entrega pipeline production-ready com capacidade de melhoria contínua.

README

A tradução para o português deste README ainda está na fila.

ai-video 🎬

The AI video skill that learns.

End-to-end video generation for Claude Code — across six models, from prompt to finished cut. It quality-controls every clip it makes, and gets better at prompting with each one.

Claude Code Models Pipeline Python


Most AI-video tooling stops at "here's your clip." ai-video doesn't. Every generation ends with a quality-control pass — Claude watches the rendered clip, scores it against the prompt, and writes down what it learned. That lesson is loaded back in before the next prompt is written. The skill is not static; it compounds.

It also covers the whole job, not just the model call: generate a first frame, write the prompt, run it on the right model, synthesize voiceover, stitch the clips, upscale for delivery.

Showcase

Real clips, real prompts — each generated with the structure this skill teaches. Open a prompt to see the 5-part craft.

[00:00-00:05] Extreme wide IMAX 70mm desert shot in the Denis Villeneuve and
Greig Fraser register: a tiny armored convoy races across hard salt flats
while a mile-high sandstorm advances behind it, desaturated ochre palette,
enormous negative space.
[00:05-00:10] Cut inside the lead rover, medium close-up of the driver
gripping the wheel at ten and two, dashboard amber reflected in the visor.
Camera shakes with engine vibration. He says, "Do not slow down." Sand
hammers the windshield.
[00:10-00:15] Exterior low tracking shot as the rover crests a dune in real
time, not slow motion. Dust swallows the frame, distant bass rumble, no gore,
no text, hard cut to black.
[00:00-00:05] Interior cockpit medium close-up of a veteran race driver at
night, rain lashing the windshield, dashboard LEDs reflected in the visor. He
tightens both gloved hands on the wheel at ten and two, breathing steady.
[00:05-00:10] Cut to rival cockpit, younger driver in the next car, jaw
tense, eyes forward, engine vibration shaking the frame. He whispers, "Hold
the line."
[00:10-00:15] Low exterior tracking shot as the green light hits and both
cars accelerate in real time on wet asphalt. Massive water spray hits the
lens, stadium lights stretch into motion blur, engines roar, no limb
distortion, no text.
Extreme macro close-up of a steak hitting a black cast-iron skillet,
background fully blurred. Fat renders, bubbles along the crust, butter foams
around rosemary and garlic, steam blooms into warm tungsten light. Slow
dolly-in, probe-lens feel, razor-thin depth of field. Audio: aggressive
sizzle, fat crackle, one tong clink at end, no music, no voice, no text.
Loop-ready studio macro shot of a perfect liquid mercury sphere on a black
mirror surface. Locked camera, centered composition, cool white strip lights
reflected across the metal. Over 6 seconds the sphere slowly deforms into a
rounded cube under an invisible force, then returns exactly to the original
sphere shape by the final frame. First and last frames match for seamless
playback. No text, no logo, no extra objects, no camera movement.

All four prompts — and 26 more — ship in examples/prompts.json.

🎞️ See the full results in motion — the complete clip library, with video previews and filterable prompts, lives at ai-video-guide.vercel.app/library.html.

Features

Install

git clone https://github.com/0xadvait/ai-video-skill.git ~/.claude/skills/ai-video
pip install -r ~/.claude/skills/ai-video/scripts/requirements.txt

ffmpeg + ffprobe must be on PATH (frame QC, stitching, upscaling). The skill activates automatically — just ask Claude Code to make a video.

Heads up: API keys are read from the environment. You only need a key for the models you actually call (see Environment).

The pipeline

A single clip needs only stages 2–4. A finished video runs the whole chain.

#StageScriptWhat it does
1Stillimagegen.pyFirst-frame / reference images — Flux Schnell, Flux Pro, Imagen 4
2PromptSKILL.md + reference/5-part structure (Subject→Action→Camera→Style→Constraints) + time-coded blocks
3Generatevalidate.py → generate.pyValidate the request, run on the auto-routed model
4QC + learnreview.py → LESSONS.mdWatch the clip, score it, record a prompting lesson
5Voiceovertts.pyElevenLabs text-to-speech narration
6Assembleassemble.pyStitch clips — cuts/crossfades + audio bed — from a JSON edit list
7Finishupscale.pyUpscale resolution + interpolate frame rate

Models

One canonical schema (Seedance 2.0's). generate.py translates it to whichever model is chosen — "model": "auto" (default) routes by intent.

ModelBackendBest for
seedance-2.0ReplicateGeneralist — quad-modal refs, native audio, lip-sync, motion transfer
kling-3.0-omnifalt2v / i2v with a different moderation gateway (the --fallback)
wan-2.7-i2vReplicateOpen-weights image-to-video
veo-3.1Google GeminiHighest-realism cinematic text-to-video
omnihuman-1.5falAudio-driven talking avatar, gesture-rich
veed-fabric-1.0falMouth-only lip-sync — included, but see LESSONS.md

The learning loop

This is the part that makes it more than a wrapper:

prompt ──▶ generate ──▶ review.py ──▶ Claude watches the clip
   ▲                                          │
   │                                          ▼
LESSONS.md ◀────────────── distill a generalizable lesson

review.py extracts a frame contact sheet; Claude scores the clip on prompt fidelity, consistency, motion, and audio; a one-line cause→effect lesson is appended to LESSONS.md. Step 0 of every workflow reads LESSONS.md back in. No regeneration is needed for the skill to improve — clips are only re-run when you ask.

Documentation

FileWhat's inside
SKILL.mdThe workflow, model routing, QC loop, hard rules
LESSONS.mdSeed lessons + the skill's growing memory
reference/schema.mdCanonical Seedance 2.0 request schema
reference/prompt-logic.md10 prompt categories, theory + perfect examples
reference/style-library.mdCross-model camera / lens / lighting / director vocabulary
reference/real-person.mdReal-person workflow + production stacks
reference/failure-modes.md15 failure modes and their fixes
examples/prompts.json30 verified, paste-ready prompts

Environment

VariableUsed by
REPLICATE_API_TOKENseedance-2.0, wan-2.7-i2v, Flux stills, ML upscale
FAL_KEYkling-3.0-omni, veed-fabric-1.0, omnihuman-1.5
GEMINI_API_KEYveo-3.1, Imagen 4 stills
ELEVENLABS_API_KEYtts.py voiceover

Personal profile

The "a video of me" shortcut uses a personal profile. Copy the template and fill in your own assets:

cp profile/profile.example.json profile/profile.json

profile.json is git-ignored — your avatar paths and voice id never leave your machine.

A note on real-person video

Real-person generation is policy-sensitive. Only build around public figures in legitimate commentary or satire, or consenting, tagged collaborators — never undisclosed impersonation of private individuals. Respect voice-provider restrictions when sourcing reference audio. See reference/real-person.md.

Contributing

Issues and PRs welcome — new verified prompts for examples/prompts.json, model adapters in scripts/generate.py, and reference-doc corrections are all fair game. Keep prompt examples real (rendered, not invented).

License

MIT © 2026 Advait Jayant.


Built from the ai-video-guide research corpus.

🎬 Generated with Claude Code

Ver o original no GitHub

Bom para

Relacionados